We develop a method for automatic segmentation of natural video sequences. The method is based on low-level spatial and temporal analyses. It features three designs to help facilitate good region segmentation while keeping the computational complexity at a reasonable level. Firstly, a preliminary seed-area identification and a final re-segmentation process are performed on each video frame to help region tracking. Secondly, a simple way to measure homogeneity of texture in a region is devised and the segmentation tries to locate object boundaries at where the texture shows significant changes. And thirdly, a reduced-complexity motion estimation technique is used, so that dense motion fields can be computed at a reasonable complexity. The overall method is organized into four tasks, namely, seed-area identification (for each frame), initial segmentation (only for the first frame in the sequence), motion-based segmentation (for all later frames), and region tracking and updating (also for all later frames). Some examples are provided to illustrate the performance of this method.
|Number of pages||9|
|Journal||Proceedings of SPIE - The International Society for Optical Engineering|
|State||Published - 1 Jan 2002|
|Event||Viual Communications and Image Processing 2002 - San Jose, CA, United States|
Duration: 21 Jan 2002 → 23 Jan 2002
- Image sequence segmentation
- Object tracking